Using AI as Decision Paths Inside your RPA Processes

Right now, I’d separate AI into two broad categories:

• Off the shelf (OTS), generic AI that doesn’t need any specialized knowledge to understand. The functions that this AI provides is something that a child could probably perform, such as categorizing/identifying things in images, reading (doing OCR), text sentiment analysis, etc. These types of AI services are provided by all of the major AI vendors like GCP and AWS.
• A customized AI solution that is application or industry specific. These types of AI can potentially replace some human expertise, such as a pharmacist’s prescription. This type of AI service will less likely be offered by the big vendors and would likely be custom built for a few companies.

In terms of integrating OTS AI, they will mostly be online (not installed on premises), requiring a connection to the Internet. This can pose a problem with security since many RPA installations are kept in an Intranet without Internet connectivity. Data security concerns also come to the forefront as your data has to be sent to a third party.

Connecting your RPA to these OTS systems would be as simple as doing an API call, where you’d first set up an account on the vendor’s site, get the required auth tokens/API keys and pass in data in their requested format.

What you would get back is a “label” and a “confidence level” (numeric between 0-1). I can see these confidence levels being used to choose between different paths in an RPA process. Let’s say that you are a veterinary office where people can upload a photo of their pet. The AI system could analyse the image and decide the “label”: is this “cat”, “dog” or “hamster”. The label would be accompanied by a confidence level, such as “cat”:0.9, “dog”:0.3 etc. meaning that the AI system is highly confident that the image is of a cat. The RPA system would have decision paths, such as “If the label is “cat” and confidence is above 0.8″ then perform X.

Customized AI would be used in a similar way, but the human function that the AI replaces would be much more complicated. For example, let’s say we have an AI system that takes in images car accidents. Let’s say that the insurance company wants to pre-fill in some forms based on the amount of damage estimated to each vehicle. The AI system could do something like: 1) identify the number of cars involved in the accident, 2) identify the make/year of the cars in the image, 3) predict the $ amount of damage done based on the photo. These are functions that an expert could probably perform, but not a normal person. But the output of the AI system could be:

[“car1”: [{“model”: “toyota camry”, “year”: “2011”, “confidence”: 0.7, “costs”: 2000}, {“model”: “toyota camry”, “year”: “2013”, “confidence”: 0.95, “costs”: 5000}, …], “car2”:[{“model”: “ford f150”, “year”: “2017”, “confidence”:0.8, “costs”: 700}…]…]

And decision paths could be designed based on the confidence levels and label of the AI system, similar to the OTS AI services.